Keshab Parhi

Keshab K. Parhi (Life Fellow, IEEE) received the Ph.D. degree in EECS from the University of California at Berkeley, Berkeley, CA, USA, in 1988. He is currently the Erwin A. Kelen Chair of Electrical Engineering and a Distinguished McKnight University Professor with the Department of Electrical and Computer Engineering, University of Minnesota Twin Cities, Minneapolis, MN, USA. He has authored or co-authored over 750 articles including 16 that have won best paper or best student paper awards, is the inventor of 36 patents, and has authored the textbook VLSI Digital Signal Processing Systems: Design and Implementation (Wiley, 1999). His current research addresses VLSI architectures for machine learning, hardware security, quantum codes, and data-driven neuroscience with focus on quantitative machine learning for neuro-psychiatric disorders.

Dr. Parhi is a fellow of the Association for Computing Machinery (ACM), the American Institute for Medical and Biological Engineering (AIMBE), the American Association for the Advancement of Science (AAAS), and is a Life Fellow of the National Academy of Inventors (NAI). He was a recipient of numerous awards including the 2003 IEEE Kiyo Tomiyasu Technical Field Award, and the 2017 Mac Van Valkenburg Award and the 2012 Charles A. Desoer Technical Achievement Award from the IEEE Circuits and Systems Society. He served as the Editor-in-Chief of the IEEE Transactions on Circuits and Systems, Part-I during 2004 and 2005, and currently serves as the Editor-in-Chief for IEEE Circuits and Systems Magazine.

Contact: parhi@umn.edu

Website: http://ece.umn.edu/~parhi

Energy-Efficient AI: From Training and Inference to Multi-Intelligence AI Agents

Abstract: The unprecedented power of large language models holds promise for design of computing systems that can achieve general intelligence by demonstrating broad capabilities of intelligence, including reasoning, planning, and the ability to learn from experience, and with these capabilities at or above human-level (Bubeck et al., GPT-4 paper). However, the path forward is unsustainable with respect to energy efficiency and growth in data required to train models. I will begin with a brief review of history of neural networks. I will talk about our prior work on Perm-DNN based on permuted-diagonal interconnections in deep convolutional neural networks and how structured sparsity can reduce energy consumption associated with memory access in these systems. I will then talk about reducing latency and memory access in accelerator architectures for training DNNs by gradient interleaving using systolic arrays. Then I will present our recent work on LayerPipe, an approach for training deep neural networks that leads to simultaneous intra-layer and inter-layer pipelining. This approach can increase processor utilization efficiency and increase speed of training without increasing communication costs. I will argue about the need for multi-intelligence AI agents at the cloud and the edge that will need to be trained using much less data by exploiting novel reasoning approaches.

Accelerator Architectures for Post-Quantum Cryptography and Homomorphic Encryption

Abstract: In this talk, I will describe accelerator architectures for post-quantum cryprtography and for homomorphic encryption for computing in the encrypted domain,. Both problems are solved using ring learning with errors (RLWE) problem. I will describe architectures based on fast time-domain and frequency-domain approaches. In the time-domain, fast filter algorithms proposed in signal processing can be adapted to implement fast lattice-based cryptosystems. For frequency domain, fast architectures can be derived using number-theoretic transforms (NTT) and inverse NTT (iNTT). I will present low-latency architectures for computing polynomial modular multiplication using NTT and iNTT. Architectural tradeoffs associated with serial and parallel architectures will be presented.

Data-Driven Neuroscience, Neurology and Psychiatry: Brain Connectivity and Classification

Large amounts of imaging data from the brain are now available to better understand the brain and “reverse engineer” the brain. Signal processing and machine learning can unravel mysteries of the brain and can be used to diagnose various brain disorders. This talk will describe analysis of functional magnetic resonance imaging (fMRI) data collected from healthy subjects at the Center for Magnetic Resonance Research (CMRR) of the University of Minnesota (UMN) as part of the U.S. Human Connectome Project (HCP), analysis of electroencephalogram (EEG) for prediction and detection of seizures from publicly available datasets, and analysis of fMRI data collected at the UMN from adolescents with psychiatric disorders and healthy controls. One goal of the analysis is to extract appropriate features and design appropriate classifiers. Sub-graph entropy, a measure of static connectivity, is introduced to discover predictive subnetworks that are used to classify task vs. no-task or to discriminate two tasks from the task fMRI data collected from the healthy subjects from HCP.

About 1% of world population suffer from epilepsy. Spectral-domain features, such as spectral powers in different bands and ratios of spectral power of two different bands extracted from EEG and intra-cranial EEG, are used to predict and detect seizures with high sensitivity and specificity. Band-power ratios of MEG during word processing task are used as features to identify subjects with schizophrenia. Resting-State fMRI data are used to design classifiers for identifying three types of psychiatric disorders among adolescents: borderline personality disorder (BPD), obsessive compulsive disorder (OCD) and major depressive disorder (MDD) using spectral-domain features and static brain connectivity. In summary, extracting appropriate biomarkers using spectral-temporal-spatial signal processing approaches and classifying states using machine learning approaches can assist clinicians in predicting and detecting various brain disorders and to understand more about the healthy brain. These biomarkers can be tracked to design personalized therapy and effectiveness of therapy by closed-loop drug delivery or closed loop neuromodulation, i.e., brain stimulation either by invasive or non-invasive means using electrical or magnetic stimulation.